GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
TheShed
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
TheShed
Model Tree
endless-frontier/BigBang-v1
Input Modalities
Output Modalities
Supported Functionality
| Component | Format |
|---|---|
| MoE and shared-expert MLP weights | W4A16 NVFP4, group size 16 |
| Attention projections | FP8 |
| Language-model output head | W4A16 NVFP4 |
| Vision and other excluded modules | BF16 |
| Native MTP module | BF16 |
The checkpoint was produced with NVIDIA ModelOpt 0.45.0 and is approximately
21.8 GiB across four safetensor shards. Exact layer assignments are recorded in
hf_quant_config.json.
Tested with vLLM 0.25.1, FlashInfer attention, the Marlin MoE backend, a BF16 KV cache, and Qwen reasoning/tool parsers.
bash
vllm serve TheShed/BigBang-v1-W4A16-NVFP4-MTP \--trust-remote-code \--kv-cache-dtype bfloat16 \--attention-backend flashinfer \--moe-backend marlin \--speculative-config '{"method":"mtp","num_speculative_tokens":1,"moe_backend":"triton"}' \--reasoning-parser qwen3 \--enable-auto-tool-choice \--tool-call-parser qwen3_xml
MTP is optional. To serve the target model without speculative decoding, omit
the --speculative-config argument. A separate non-MTP checkpoint is not
required.
endless-frontier/BigBang-v1Qwen/Qwen3.6-35B-A3BReleased under the Apache License 2.0, matching the upstream model.